Out of distribution detection

Identifying inputs that differ from the data a system was developed or evaluated on.

Unfamiliar inputs can warrant additional review, but novelty and error are different concepts. A new scene may be perfectly valid, and a familiar scene may contain a wrong label. Confidence scores alone do not reliably establish whether an example lies outside a model's experience.

In practice

A new camera angle prompts a coverage check rather than an automatic decision to discard the recording.